On Wed, 2025-04-30 at 09:38 +1000, Ben Skeggs wrote:
> Patch 1 is a simple fix to the condition used to wait for RM shutdown.
> It's not known to fix any issue, but kept separate in case it needs to
> be backported.
>
> Patches 2-4 bump up the maximum instance counts of various engines to
> support those available on r570 / newer HW.
>
> Patches 5-13 split up the GSP-RM code into modules (roughly) around RM
> API boundaries, and move all the r535-specific code and headers under
> nvkm/subdev/gsp/rm/r535 to make it easier to contain version-specific
> code.
>
> Patches 14-18 reduce the amount of boilerplate needed to implement
> engines and engine objects when running on GSP-RM by adding a common
> implementation which can be used for all engine types.
>
> Patches 19-20 switch to using the hInternal{Client,Device,Subdevice}
> objects returned by GSP during init instead of allocating them for
> each nvkm client, reducing the number of RPCs needed per FD open, and
> removing the need to handle changes in r570.
>
> Patches 21-44 add a bunch of simple HALs around the RMAPI calls that
> change between r535 and r570, and, finally, adds support for 570.144.
>
> Patches 45-52 add support for GH100 Hopper GPUs
>
> Patches 53-56 add support for GB10x Blackwell GPUs
>
> Patches 57-60 add support for GB20x Blackwell GPUs
Reviewed-by: Timur Tabi <tt...@nvidia.com>
I only glossed over the DRM-specific code, since I'm not familiar with it.
However, I do have one
general complaint. Except for a line or two here and there, none of the code
is documented. I
would love to see some patches posted in the future that addressed this.
Patches 52 and 60 are missing from this posting. I do have comments on patch
52, which will
hopefully show up soon.
